As training proceeds, Gaussians get deleted and the map has holes in it.

[projektjoe]

I have tested both omni-re and street gaussian implementations within drivestudio. As the number of iterations increase, I noticed that the file size of the checkpoint decreases

I have also tried to view the scene in a 3D viewer (birdeye-view) I noticed that there are a lot of black holes (50k iter):

For 100k, it gets significantly worse.

I did not change any hyperparameter, except for the number of iterations.

[Wzxthu]

After 20000 iter, it only did culling but not duping. That's why it is getting smaller I think.

[missTL]

How should I export PLY files？Thanks

[projektjoe]

How should I export PLY files？Thanks

check this #26

[projektjoe]

After 20000 iter, it only did culling but not duping. That's why it is getting smaller I think.

I start to see that trend before 20k. Potentially at 15k or even before it. I also do not expect this to happen, since with all other 3DGS methods, you can increase iterations without quality degradation.

[Wzxthu]

Actually I also noticed that issue when I digged in more cases lol

[ziyc]

Hi, @projektjoe, our hyperparameters are designed for standard 30k iterations of gaussian training. Under this setting, densification stops after 15k iterations while the pruning process continues. When training for longer periods (e.g. 100k iterations), your gaussians will start to decrease after 15k as points continue to be deleted, that's why your file size keeps shrinking and holes appear in the scene.

For longer training iterations, please consider tuning essential hyperparameters accordingly, especially the densification settings here:

drivestudio/configs/omnire.yaml

Line 67 in 388642a

stop_split_at: 15000 # stop splitting at this step

[projektjoe]

I tried to make stop_split_at = total iterations and it did indeed produce a map with no holes

1. Initially, I was doing 30k total iterations and stop_split_at 15k and the map had the holes, is there a way to adjust the hyperparameter stop_split_at properly? since I believe it shouldn't be equal to the total iterations.

2. Also, I noticed the 3D scene quality is close but not as clean as street gaussian's official implementation. Is there a way to improve the map quality even if that requires more compute?

[xiaokezhou]

@projektjoe Hi, did you figure this out? I'm frustrated with the PLY file exported via the method in #26 you mentioned. I put the ply file into the online viewer https://playcanvas.com/supersplat/editor/, and got 3d rendering result similar with the image you pasted. The videos and images rendered during the training are good as expected.